In his latest post Ulf Wendel looks at the performance issues around the recently introduced MySQL Fabric support included in the mysqlnd extension.
PECL/mysqlnd_ms 1.6 is currently being modified to support sharding and fully automatic server and client failover when using MySQL Fabric (slides) to manage a farm of MySQL servers. PECL/mysqlnd_ms is a mostly transparent load balancer that works with all PHP MySQL APIs (PDO_MySQL, mysqli, ...). The idea is, that if, for example, a MySQL server fails, the plugin talks to MySQL Fabric to learn about alternative servers that Fabric has provisioned automatically. This "talks to" gives implies a performance penalty for applications.
He takes a look at what's happening "behind the scenes" when it comes to using the Fabric functionality and sharding (based on the use of mysqlnd_ms_select_shard). He traces through the execution path and how much slower then end result is. He includes some results from the connection debugging and the number of queries a single request makes.